From c94bac43e5779bd6802bcec76800ef4f64bacf5a Mon Sep 17 00:00:00 2001
From: dimgigov <gigov@gigov.org>
Date: Fri, 31 Jul 2026 03:02:32 +0300
Subject: [PATCH] fix(raft): ignore intermediate InstallSnapshot chunk replies

---
 src/barabadb/core/raft.nim |  8 +++++++-
 tests/test_all.nim         | 30 ++++++++++++++++++++++++++++++
 2 files changed, 37 insertions(+), 1 deletion(-)

diff --git a/src/barabadb/core/raft.nim b/src/barabadb/core/raft.nim
index b96674d..7f1119d 100644
--- a/src/barabadb/core/raft.nim
+++ b/src/barabadb/core/raft.nim
@@ -679,7 +679,13 @@ proc handleInstallSnapshotReply*(node: RaftNode, peerId: string,
   if node.state != rsLeader:
     return
 
-  if reply.success:
+  if reply.success and reply.matchIdx >= node.lastSnapshotIndex:
+    # The follower has actually adopted the snapshot base. Intermediate chunk
+    # replies (the T8 follower acks every non-done chunk with success=true and
+    # matchIdx = its OLD lastSnapshotIndex, below ours) fall through here and
+    # must be ignored: applying them would regress matchIndex/nextIndex and
+    # clear the reject streak mid-transfer, causing state flapping until the
+    # final reply lands.
     node.matchIndex[peerId] = reply.matchIdx
     node.nextIndex[peerId] = reply.matchIdx + 1
     node.snapRejectStreak.del(peerId)
diff --git a/tests/test_all.nim b/tests/test_all.nim
index 29f04a9..5d758c5 100644
--- a/tests/test_all.nim
+++ b/tests/test_all.nim
@@ -2931,6 +2931,36 @@ suite "Raft InstallSnapshot Send":
     check node.nextIndex["peer-1"] == 101
     check node.snapRejectStreak["peer-1"] == 2
 
+  test "intermediate chunk replies are ignored; final reply advances state":
+    var node = newRaftNode("leader", @["peer-1"])
+    node.currentTerm = 5
+    node.state = rsLeader
+    node.lastSnapshotIndex = 100
+    node.lastSnapshotTerm = 4
+    node.nextIndex["peer-1"] = 101
+    node.matchIndex["peer-1"] = 0
+    node.snapRejectStreak["peer-1"] = 2
+    node.snapPending.incl("peer-1")
+
+    # Intermediate chunk ack: follower replies success=true with its OLD
+    # lastSnapshotIndex (40 < our 100). Leader state must not move.
+    node.handleInstallSnapshotReply("peer-1", RaftMessage(
+      kind: rmkInstallSnapshotReply, term: 5, senderId: "peer-1",
+      success: true, matchIdx: 40))
+    check node.matchIndex["peer-1"] == 0
+    check node.nextIndex["peer-1"] == 101
+    check node.snapRejectStreak["peer-1"] == 2
+    check "peer-1" in node.snapPending
+
+    # Final reply: follower adopted the snapshot base (matchIdx == 100).
+    node.handleInstallSnapshotReply("peer-1", RaftMessage(
+      kind: rmkInstallSnapshotReply, term: 5, senderId: "peer-1",
+      success: true, matchIdx: 100))
+    check node.matchIndex["peer-1"] == 100
+    check node.nextIndex["peer-1"] == 101
+    check "peer-1" notin node.snapRejectStreak
+    check "peer-1" notin node.snapPending
+
   test "InstallSnapshotReply term handling matches AppendEntriesReply":
     var node = newRaftNode("leader", @["peer-1"])
     node.currentTerm = 5